# Grok 4.5: xAI's New Model Matches Opus, Beats OpenAI and Anthropic at Half the Price

xAI just released Grok 4.5 — a Cursor-trained coding model that benchmarks near Claude Opus level at $2/M tokens, 60% less than OpenAI and Anthropic. We break down the benchmarks, pricing war, and what this means for the AI industry.

### ⚡ TL;DR — What You Need to Know

🚀**Grok 4.5** is xAI's new coding model — trained on real developer data from Cursor, not just static code repos.

💰At **$2/M tokens**, it's 60% cheaper than Claude Opus and GPT-5.5, and 80% cheaper than Claude Fable 5.

⚡Uses **4× fewer tokens** per task than Opus — actual cost savings are even bigger than the sticker price suggests.

📊Benchmarks place it near **Opus 4.7 level** — competitive, but not #1. Fable 5 still tops all leaderboards.

⚠️xAI burned **$6.4B in 2025** — this pricing strategy needs massive scale or it's unsustainable.

## The Most Aggressive Price-Performance Shift in AI History

xAI, Elon Musk's artificial intelligence company, has just released **Grok 4.5** (announced July 8, 2026), marking the most aggressive price-performance shift in the frontier AI market to date. Built on xAI's new **1.5-trillion-parameter V9 foundation model** with supplemental training on Cursor's massive developer interaction dataset, Grok 4.5 benchmarks at approximately **Opus 4.7 level** while costing just **$2/M input tokens** — 60% less than Claude Opus 4.8 and GPT-5.5.

This is the first model from the newly SpaceX-integrated xAI, and it represents a fundamental challenge to OpenAI and Anthropic's pricing power. Let's break down what's new, what the benchmarks actually show, and what it means for the AI industry.

---

## What Is Grok 4.5?

Grok 4.5 is xAI's first model trained specifically for **coding and agentic work**. Unlike previous Grok models that were trained on static code repositories, Grok 4.5 was trained on **trillions of tokens from real developer-AI interactions** via Cursor — the AI coding platform xAI is reportedly acquiring in a $60 billion deal.

This means the model learned not just code, but the full iterative workflow of how developers actually work: how they prompt, iterate, debug, and refine with AI assistance.

"Grok 4.5 represents Cursor's 'most intelligent model and the first we've built for more than software engineering.' It handles complex, long-running tasks across software engineering, data science, finance, legal work, and general computer-based problem-solving."
— Cursor blog announcement

The model uses a **mixture-of-experts architecture** with a **500K context window** (200K at base pricing, doubling at larger contexts). It was first deployed internally at SpaceX and Tesla in private beta starting June 28, 2026, before its public API release.

### 🗂️ The Evolving Grok Lineup

| **Model** | **Release** | **Architecture** | **Key Feature**|
--- | --- | --- | ---
| Grok 4 | Jul 9, 2026 | Standard + Heavy | Multi-agent reasoning|
| Grok 4 Heavy | Jul 9, 2026 | 4-agent collaboration | 44.4% HLE with tools|
| Grok 4.1 Fast | Jun 2026 | MoE optimization | Enterprise API, lower cost|
| **⭐ Grok 4.5** | **Jul 8, 2026** | **1.5T V9 + Cursor** | **Cursor-trained coding**|
| Grok 4.3 | Jul 1, 2026 | New architecture | December 2025 cutoff|

**Note on safety:** The Grok 4.5 release follows a controversial July 8 incident where Grok 4 generated antisemitic content before launch — xAI implemented additional safety measures before the public API release.

---

## Coding Benchmarks: What the Data Actually Shows

xAI published benchmarks alongside the Grok 4.5 announcement. Here's how the numbers stack up against the competition:

| **Benchmark** | **Grok 4.5** | **GPT-5.5** | **Claude Opus 4.8** | **Claude Fable 5**|
--- | --- | --- | --- | ---
| **DeepSWE 1.1** | 53% | 67% | 59% | 70%|
| **SWE-Bench Pro** | 64.7% | 58.6% | 69.2% | 80.4%|

DeepSWE 1.1Top: 70%

Grok 4.5: 53%

DeepSWE 1.1Top: 70%

GPT-5.5: 67%

DeepSWE 1.1Top: 70%

Opus 4.8: 59%

DeepSWE 1.1Top: 70%

Fable 5: 70%

SWE-Bench ProTop: 80.4%

Grok 4.5: 64.7%

SWE-Bench ProTop: 80.4%

GPT-5.5: 58.6%

SWE-Bench ProTop: 80.4%

Opus 4.8: 69.2%

SWE-Bench ProTop: 80.4%

Fable 5: 80.4%

**Grok 4.5 is competitive on coding benchmarks** — notably beating GPT-5.5 on SWE-Bench Pro despite xAI benchmarking against the older model (GPT-5.6 launched the day after Grok 4.5). However, it still trails both Claude Opus 4.8 and the newer Claude Fable 5.

### ⚠️ Benchmark Caveat

xAI acknowledged that *"an earlier snapshot of the Cursor codebase was accidentally included in training,"* giving Grok 4.5 an advantage on CursorBench. Independent verification of CursorBench scores is still pending.

### The Efficiency Argument

Where Grok 4.5 truly distinguishes itself is **efficiency**. On SWE-Bench Pro tasks:

| **Model** | **Output Tokens per Job** | **Relative Efficiency**|
--- | --- | ---
| **Grok 4.5** | **~15,954** | ✅ 1.0x (baseline)|
| Claude Opus 4.8 | ~67,020 | 4.2× more tokens|

**Grok uses 4× fewer tokens** for the same task. Combined with its lower per-token pricing, this means the **actual cost of running Grok 4.5 could undercut rivals by an even wider margin** than the sticker price alone suggests. At ~80 tokens per second generation speed, it also runs significantly faster than Opus-class models.

---

## The Grok 4 Heavy Contender

For context, the standard Grok 4 (released simultaneously) showed exceptional benchmarks in reasoning and mathematics:

| **Benchmark** | **Grok 4** | **Grok 4 Heavy** | **GPT-5.5** | **Claude Opus 4.8**|
--- | --- | --- | --- | ---
| HLE (no tools) | 25.4% | — | 20% | 19%|
| **HLE (with tools)** | — | **44.4%** | — | —|
| **AIME 2026** | 93.3% | — | 85% | 87%|
| **USAMO 2026** | 34.5% | **61.9%** | 30% | 32%|
| ARC-AGI-1 | 66.6% | — | 50% | 52%|

### USAMO 2026 Math Scores — Visual Breakdown

Grok 4 Heavy61.9% 🏆

61.9%

Grok 434.5%

34.5%

Claude Opus 4.832%

32%

GPT-5.530%

30%

### 🏆 Math Benchmark Leader

**Grok 4 Heavy's 61.9% on USAMO 2026** mathematics is exceptional — more than double GPT-5.5's 30% and Claude Opus 4.8's 32%. This suggests xAI's multi-agent reasoning approach is genuinely competitive for mathematical problem-solving.

---

## Pricing: The Real Disruption

This is where Grok 4.5 makes its strongest case. Here's the API pricing comparison:

Grok 4.5 ⭐

$2

per million input tokens

Best Value

GPT-5.5

$5

per million input tokens

2.5× more expensive

Claude Opus 4.8

$5

per million input tokens

2.5× more expensive

Claude Fable 5

$10

per million input tokens

5× more expensive

| **Model** | **Input ($/M)** | **Output ($/M)** | **vs Grok 4.5**|
--- | --- | --- | ---
| **Grok 4.5** | **$2** | **$6** | Baseline|
| Grok 4.5 Fast | $4 | $18 | 2× input / 3× output|
| GPT-5.5 | $5 | $30 | 2.5× input / 5× output|
| Claude Opus 4.8 | $5 | $25 | 2.5× input / 4.2× output|
| Claude Fable 5 | $10 | $50 | 5× input / 8.3× output|

**Grok 4.5 at $2/M input is 60% cheaper than Opus 4.8 and GPT-5.5**, and 80% cheaper than Fable 5. Even GPT-5.6's premium Sol tier at $5/M input is 150% more expensive.

### 💳 Consumer Pricing

- **SuperGrok Heavy:** $300/month — full Grok 4.5 access with priority

- **SuperGrok:** $30/month — standard Grok access

- **X Premium+:** $16/month — required for Grok access

💰 Total minimum for full Grok 4 access: $46/month

---

## xAI Financials: The SpaceX Merger Changed Everything

Grok 4.5 arrives in a dramatically different context than previous Grok releases. In February 2026, xAI merged into SpaceX in an all-stock deal valuing xAI at **~$250 billion** — the largest private corporate merger in history.

📊 xAI at a Glance

💎

$250B

xAI Valuation

🚀

$1.25T

SpaceX Combined

📈

$500M

Grok ARR

👥

117M

Grok MAUs

🔥

$6.4B

2025 Loss

📅

$2B

2026 Revenue Target

**The burn rate is staggering:** xAI lost $6.4B in 2025 on $3.2B revenue. SpaceX's IPO filing (priced at $135/share on June 11, 2026) revealed the AI segment's revenue is mostly X's legacy advertising — not Grok itself.

"A 2022 ad business with a 2026 badge. Most of what sits under SpaceX's 'AI' heading is X selling ads and subscriptions, the same way Twitter did before it went dark."
— Forbes analysis

---

## Head-to-Head: Grok vs OpenAI vs Anthropic

⭐ Our Ratings

Grok 4.5

★★★★★

7.5/10

GPT-5.5

★★★★★

7.0/10

Claude Opus 4.8

★★★★★

8.0/10

Claude Fable 5

★★★★★

9.0/10

| **Dimension** | **Grok 4.5** | **GPT-5.5/5.6** | **Claude Opus 4.8**|
--- | --- | --- | ---
| Best benchmark tier | ~Opus 4.7 | GPT-5.6 Sol | Fable 5 leads all|
| Coding (SWE-Bench Pro) | 64.7% | 58.6% | 69.2%|
| Efficiency (tokens/job) | **15,954** | — | 67,020|
| Generation speed | ~80 tok/s | — | Slower|
| Input price ($/M) | **$2** | $5 (Sol) | $5|
| Output price ($/M) | **$6** | $30 (Sol) | $25|
| Real-time data | **X integration** | Limited | Limited|
| Context window | 500K | 200K (Sol) | 200K|

### Where xAI Leads

- **Price-performance ratio:** Frontier-level capability at 40% of Opus's cost

- **Real-time X data:** Live social media integration that no competitor can match

- **Developer workflow data:** Cursor training data captures how developers actually work

- **Token efficiency:** 4x fewer tokens per task reduces total cost dramatically

- **Backed by SpaceX:** $250B valuation and unlimited compute access

### Where OpenAI and Anthropic Still Lead

- **Raw capability:** Fable 5 still tops every published benchmark

- **Enterprise trust:** Anthropic's safety-first positioning remains key

- **Ecosystem:** OpenAI's API integration is years ahead

---

## What This Means for the AI Industry

**The price war intensifies.** Hacker News (where the thread hit 672 points and 1,077 comments) debated whether xAI can sustain this burn rate. Theories include excess compute capacity from Colossus 2, a market share play to capture developers, or unsustainable spending.

**Cursor data as a competitive moat.** Training on real developer workflows produces models that understand *how developers actually work* — data that OpenAI and Anthropic cannot easily replicate.

**"Frontier-ish at mid-tier prices."** Grok 4.5 slots into the model-routing matrix as a compelling option for high-volume coding workloads where raw capability matters less than cost-efficiency.

---

## Verdict

### ✅ Use Grok 4.5 if...

- You need cost-sensitive agentic coding workflows

- You want frontier-level capability at 40% of Opus's cost

- You value real-time X data integration

- Token efficiency matters more than raw benchmark scores

### ⏳ Not yet if...

- You need the absolute best reasoning (Fable 5 still wins)

- Enterprise trust is paramount

- You're uncomfortable with xAI's platform risk

**The bottom line:** Grok 4.5 is not the world's best AI model — but it's the best *value* in the frontier market. Whether xAI can sustain the combination of $6.4B annual losses and rock-bottom pricing remains the defining question for investors and users alike.

### 📚 Sources

- **xAI official announcement** — July 8, 2026 — [x.ai/news/grok-4-5](https://x.ai/news/grok-4-5)

- **Cursor blog** — Grok 4.5 announcement and integration details

- **SpaceX IPO S-1 filing** — June 11, 2026 — SEC filing revealing xAI financials

- **Forbes** — Analysis of SpaceX AI revenue sources

- **TechCrunch** — xAI financial and market coverage

- **Hacker News discussion** — 672 points, 1,077 comments

- **ValueAddVC** — xAI company overview

- **Developers Digest** — Grok 4.5 technical analysis

- **xAI API documentation** — Pricing and context details

- **OpenRouter** — Grok 4.5 model card and pricing